Hi everyone,

I am dividing visual stimulation (task) betas by breath-hold betas (bh; 2 separate scans) to reduce cerebrovascular contributions to the task BOLD signal. Some of the bh betas are very low and so division results in "interestingly" large task betas. These greatly distort pre/post-scaling investigations.

Question: What are methods/decision rules to eliminate extreme values when standard deviation around the mean is not helpful?

Example: after division, a cluster has 606 voxels with scaled task betas ranging from -1.9 -> 8.7 ... and then there are these: -11, -15, -22, -38, 19, 154 (cluster mean=1.59, sd= 6.66, median=1.11).

I'm currently looking into median absolute deviation (MAD; assume normal distribution) and have set an arbitrary value of 7 MAD. This "far out there" MAD is catching the extreme values (and sometimes a few more that I didn't think of as extreme but this is how it is) and so far seems to work.

Is this a problematic/biased approach to pre/post-scaling investigations? I'm a bit at a loss for an appropriate method/decision rule on how to remove these extreme values.

Any suggestion is greatly appreciated!
Thank you very much in advance,

​Nic​


​-------------------------- More Information If Desired --------------------------

Analysis: FWE cluster-thresholding of a one-sample t-test of visual stim betas provides a binary mask to extract the visual stim and bh betas at the single-sub level. In each participant and on a voxel-wise basis, scaling is carried out by dividing the visual stim betas by the bh betas in the clusters of this mask.

Various efforts to counter the extreme values: using tent functions in the single-sub GLM & grey matter masks during the scaling operation to get "appropriate" bh betas in the first place to scaling by dividing by a) simple bh beta, b) normalized bh beta, c) normalized bh beta +1, and d) not including bh betas between +/- 0.1.

There is always a combination of visual stim & low bh betas that gives these large values. As such, it appears to be time to concentrate on the output, rather than the input.